negative samping based on popularity

[yizerozhuang]

For negative sampling and add_negative_history in DIEN procssing Recdp and ai matrix generate a list that contains all the items in the reviews including duplicate items and randomly select a negative sample from the list. Friesian encodes the items first, then randomly select one of the items within the largest index range as negative example.

So while adding negative examples, ai-matrix and recdp tends to select the items that have higher frequencies while Friesian has equal chance for every item.